Improve token estimation accuracy for multibyte text

[bug-ops]

Parent: #740 (P0)

Problem

Current estimate_tokens() uses bytes/3 heuristic which overestimates on multibyte text (Cyrillic, CJK — up to 2-3x). This causes premature context compaction and inaccurate budget allocation.

Solution

• Replace text.len() / 3 with text.chars().count() / 4
• Add configurable safety margin (default 1.0, recommended 1.2 for production)
• Optionally support tiktoken-rs behind a feature flag for precise counting with cloud providers

Affected crates

• zeph-memory (estimate_tokens function)
• zeph-core (all call sites)

Acceptance criteria

• Estimation accuracy within 20% on mixed ASCII/Cyrillic/CJK text
• Safety margin configurable via memory.token_safety_margin
• Existing tests updated